Development of a Hybrid XGBoost and CNN-based Deep Learning Model for Diabetes Prediction

Diabetes is a chronic disorder, and early detection is critical. Traditional machine-learning models—logistic regression, decision trees—often miss the complex, nonlinear interactions among clinical variables. The hybrid XG BoostCNN architecture that unites gradient-boosted feature selection with deep pattern recognition to enhance predictive accuracy on open-source Kaggle Pima-Indian dataset. After imputing missing values, scaling features, and winsorizing outliers, we retain eight routine predictors: Glucose, BMI, Blood Pressure, Insulin, Skin Thickness, Pregnancies, Diabetes Pedigree, and Age. XG Boost first ranks feature importance; a compact onedimensional CNN then learns discriminative representations from the selected features. Trained with the Adam optimizer and binary cross-entropy loss since it has speed and adaptability during training and binary cross entropy is the best choise when binary classification is used like predicting whether a patient has disease or not. The hybrid achieves 90.5 % accuracy— substantially higher than logistic regression (78.5 %), decision tree (79.2 %), standalone XG Boost (85.3 %), and standalone CNN (86.7 %). No other evaluated of classifier surpass its performance metrics. The results underscore the model’s potential for robust, scalable, data-driven diabetes screening in clinical settings.